从医学院到 AI：一位生物学家的旅程

Daria，非常感谢你来到这里。

Daria, thank you so much for being here.

谢谢。我非常兴奋。

Thank you. I'm very excited.

同样。非常高兴能邀请到你，因为你显然是一位非常独特的构建者，与我们通常交谈的构建者截然不同。你有医学背景，但也从事生物科学和生物工程工作，并且以一种大多数构建者不会的方式推动 AI，带着你在生物学、癌症、免疫学等众多主题上的深厚底蕴。所以非常期待深入探讨。

Likewise. Very excited to have you because you're obviously a very unique kind of builder, quite unlike the builders we usually talk to. You have a medical background but you're also working in bioscience and bioengineering, and you're pushing AI in a way that most builders don't, coming at it with this real depth that you have in so many topics from biology to cancer to immunology. So very excited to dive in.

谢谢。非常感谢。也许第一个问题是：如果我们回溯，你是什么时候第一次意识到生物学和科学将需要 AI？

Thank you. Thank you very much. Maybe the first question: if we rewind, when did you first realize that biology and science were going to need AI?

是的，我想那是在我医学院毕业后，当我意识到生物系统的复杂性时。事实上，毕业后我进入了生物医学研究，因为我真的想理解生物学，我发现在当时我们无法治疗很多疾病，就是因为复杂性。我越深入研究，就越意识到：“天哪，这太不可思议了。我们该如何解决这个问题？”生物系统中有数万亿个不同的部分和数十亿个反应在发生，这简直让人不知所措。那是我在 90 年代初对 AI（人工智能）产生兴趣的时候。然后我意识到也许有一天我们可以用 AI 构建模型，并在 90 年代非常热衷于尝试用 AI 编程。当然，快进到深度学习革命，我非常兴奋，因为第一次我看到深度学习能够以并行方式处理海量信息，当然还有 AlphaFold 和后来的 ChatGPT。但最初的时刻是我医学院毕业的时候。

Yeah, I think that was right after I graduated from medical school, when I realized the complexity of the biological system. In fact, after I graduated I went into biomedical research because I really wanted to understand the biology, and I saw that at the time we couldn't treat a lot of diseases because of the complexity. The more I went into the research aspect, the more I realized, 'Oh my god, this is incredible. How are we going to solve this?' There are trillions of different parts in biological systems and billions of reactions happening, and it just seemed so overwhelming. That's when I got interested in AI, artificial intelligence, in the early 90s. Then I realized maybe someday we could actually build models using AI, and I got very interested in trying to code with AI during the 90s. Of course, fast forward to the deep learning revolution, I was so excited because for the first time I could see that deep learning is able to process this massive amount of information in a parallel fashion, and of course AlphaFold and then ChatGPT. But the first moment was when I finished my medical school.

太棒了。自从 ChatGPT 发布以来，感觉你一直活跃在我们的社区里，对吧？你一直在测试模型。我记得你用 o1 preview 做过一些工作。当我们有了第一批推理模型时，你第一次上手时的反应是什么？

Amazing. And since that moment of launching ChatGPT, it feels like you've very much been in our community, right? You've been testing models. I remember you doing some work with o1 preview. When we had the very first reasoning models, what was your first reaction when you started to get your hands on that?

是的。我仍然记得那是 2024 年 9 月。实际上，OpenAI 联系了我，我想是因为我在 X（以前的 Twitter）上非常活跃，谈论 AI 以及它将如何改变人类。当时有很多怀疑的人，我想现在仍然有，但我真的相信它，并全力投入 AI。我想 OpenAI 对我感兴趣，让我尝试第一个推理模型。我仍然记得我给了它一个关于免疫学的非常复杂的问题。但那个问题，我甚至还记得提示词。我对游戏、电子游戏非常感兴趣，所以我喜欢把游戏交叉虚拟化到科学中。我玩的一个游戏是一种生存游戏，你在一个岛上战斗。我不会提公司的名字，但你明白那种游戏类型。这是一种大逃杀游戏。在某种程度上，免疫系统就像是对抗肿瘤的大逃杀。我说：“想象一下大逃杀游戏和免疫系统。你会如何设计一个免疫细胞对抗癌症的场景？”这是一个完全不同的领域。事实上，我们后来确实做了相关的实验。而 o1 preview 给我的回答几乎让我激动不已。我当时想：“天哪。”在那之前，GPT-4 不会给你那种深刻、有洞察力的回答。我仍然记得那一天。那非常特别。

Yes. I still remember this was September 2024. Actually, OpenAI approached me, I guess because I was very active on X (or Twitter in the old days), talking about AI and how this was going to change humanity. At the time there were a lot of skeptical people, and I think there still are, but I truly believed in it and went all in on AI. I guess OpenAI was interested for me to try, I guess, the first reasoning model. I still remember when I gave it a very complex question on immunology in my topic. But the question, actually I even remember the prompt. I'm very interested in games, video games, so I like to sort of cross-virtualize games to science. One of the games I would play is a sort of survival game where you're on an island and you're fighting. I won't give the names of the companies, but you understand the type of games. It's kind of a battle royale game. In a way, the immune system is kind of a battle royale against tumors. I said, 'Think of battle royale games and the immune system. How would you develop a scenario where the immune cells would fight against cancer?' It's a completely different field. In fact, we actually did experiments related to that later on. And o1 preview just gave me this response that almost made me emotional. I was like, 'Oh my god.' Before that, GPT-4 wouldn't really give you that kind of deep, insightful response. I still remember that day. That was very special.

你会说那是你认为“好吧，AI 现在对科学来说是不可避免的”的时刻吗？

Was that the moment you'd say where you thought, 'Okay, AI is now inevitable for science'?

绝对是的。我的意思是，即使在那之前，尤其是在 GPT-4 发布之后，它已经非常有用。我会告诉我的同事们，因为信息太多了，尤其是在生物学领域，你根本跟不上。对于搜索文献和真正综合知识，当然，还有做像写信这样的琐事——我写很多推荐信——以前需要一个小时的事情现在只需要五分钟。但还没有到你可以完全信任它或问诸如“实验的结果会是什么？”这样的问题的程度。我认为 o1 preview 是那个时刻，我意识到当它开始推理时，你得到的东西将对科学极其相关。当然，在那之后我们有了 pro 版本和 o3，它们变得更好，现在我们有了令人难以置信的模型。我可以提到我目前的体验，它正在变得越来越好。

Absolutely. I mean, even before that, especially after GPT-4 came out, it was extremely useful. I would tell my colleagues, because there's so much information, especially in biology, you can't really keep up with it. For searching literature and really synthesizing the knowledge, of course, doing mundane things like writing letters—I write a lot of letters of recommendation—things that used to take an hour would take five minutes. But it wasn't at the point where you could fully trust it or ask questions like, 'What would be the outcome of an experiment?' for example. I think o1 preview was that moment where I realized when it starts to reason, then you get things that are going to be extremely relevant for science. Of course, after that we had the pro version and o3, which got better, and now we have unbelievable models. I can mention my current experience, which is getting better and better.

是的。也许说到当下，快进到今天。几个月前我读到你的推文，关于你的新日常：现在你开始一天时，先是早上的咖啡，然后 Claude Codex 必须为你做点什么。

Yeah. Maybe speaking of the current moment, fast forward to today. I read a couple months ago your tweet about your new routine where you can start your days now with first a morning coffee but second, Claude Codex has to be doing something for you.

是的。所以我可以说自己是一个 Codex 成瘾者。我绝对是一个认证的成瘾者。早上醒来第一件事，因为 Codex 实现的是，有各种各样的想法——无论你想做一个模拟，还是想做一个特定的应用，或者想做一个游戏，或者其他什么。在过去，你必须知道如何编程。即使你知道如何编程，也需要几周或几个月的时间。但现在，我有很多想法，醒来后，喝杯咖啡，我说：“好吧，我得试试这个想法。这会怎么样？”有时 Codex 实际上在通宵工作，所以我想看看结果是什么。过去几个月我睡得不多，因为太着迷了。所以是的，我爱 Codex。

Yes. So I can call myself a Codex addict. I'm definitely a certified one. The first thing in the morning when I wake up, because what Codex enabled is, there are all these ideas—whether you want to make a simulation, whether you want to make a certain app, or you want to make a game, or whatever it is. In the past, you had to know how to code. Even if you knew how to code, it would take you weeks or months to do it. But now, I get a lot of ideas, and the moment I wake up, get my coffee, I say, 'Okay, I got to try this idea. How is this going to work out?' Sometimes actually Codex is working overnight, so I want to see what the results are. I haven't slept that much in the last couple of months because of that fascination. So yeah, I love Codex.

我很好奇，对于像你这样在免疫学、肿瘤学、癌症、T 细胞方面有深厚底蕴的人，你是如何使用 Codex 将这些世界结合起来，并且对你的日常也有用的？

I'm curious, for someone like you with real depth in immunology, oncology, cancer, T cells, how do you approach using Codex to bring these worlds together in a way that's also useful to your day-to-day?

是的。我甚至可以展示具体的例子。我一直在尝试构建一些不是非常复杂但日常极其有用的应用。我有很多这样的例子。但此外，我们非常依赖软件进行分析。正如我所说，生物学非常复杂，无论是遗传学。

Yeah. I can even show specific examples. I've been trying to build some apps that are not very complex but extremely useful on a day-to-day basis. I have a bunch of examples of that. But also, we are very much dependent on software for analysis. As I said, biology is very complex, whether it's genetics.

在免疫学中，我们经常做一种叫做流式细胞术的分析。基本上，这就像我们观察细胞世界的窗口，主要是免疫细胞，但你可以分析任何细胞。我们有这些带大激光的机器。我们用特定的标记物——荧光标记物——来标记细胞，因为细胞有数百种不同类型。那么，你怎么知道血液或组织里有什么类型的细胞呢？我们标记它们，让它们通过激光，激光分析数千个细胞，然后生成数据，告诉我这是一个能对抗癌症的免疫细胞，那个可能引起自身免疫，或者别的功能。但我们必须把这些数据导入专门的软件。软件处理成千上万甚至几十万个单个细胞的数据点，生成图表，这样你就可以说：‘给我这些细胞的百分比，以及它们之间的关系’等等。这是我们过去几十年一直在用的非常复杂的软件。有一天我说：‘为什么不自己做一个呢？’这是个疯狂的想法，因为非常复杂，实际上我失败了很多次。我尝试了，有些东西成功了。但自从 5.5 版本以来，我现在有了一个完全可用的版本。

In immunology, for example, we do a lot of analysis called flow cytometry. Basically, this is like our window to the world of cells, mainly immune cells, but you can analyze any cells. We have these machines with a big laser. We label the cells with particular markers, fluorescent markers, because there are hundreds of different types of cells. So, how do you know what type of cells are in your blood or tissues? We label them, pass them through this laser, and the laser analyzes thousands of cells, then creates data that tells me this is an immune cell that can fight cancer, this one can cause autoimmunity, or whatever the function is. But we have to take that data and put it into specialized software. It takes all those thousands or sometimes hundreds of thousands of data points from individual cells, and the software makes graphs so you can say, 'Give me the percentage of these cells there and how they relate to that,' and so on. This is very sophisticated software we've been using for the last couple of decades. One day I said, 'Why not make one for myself?' It was a crazy idea because it's very complex, and in fact, I failed a lot. I tried, some things worked. But since 5.5, I now have a completely working version of this.

太棒了。如果你笔记本电脑上有东西可以展示，我很想看看。

That's amazing. I would love to see it if you have things on your laptop to show me.

我可以给你看一些例子。从这个开始。我有不同的版本，但我会展示其中一个应用。基本上，我已经上传了一个文件，你可以看到这些点每一个都是一个细胞。

I can show you some examples. Start with this. I have different versions, but I'll show you one application. Basically, I already uploaded a file, and you can see that each of these dots is a single cell.

嗯。

Mhm.

这些是荧光分子的颜色，如果你愿意这么理解的话。任何带有那种荧光分子的抗体会标记特定的细胞类型。我可以在这里选择：‘改变我’。你可以看到所有这些都是用不同抗体标记的，非常复杂，因为有 20 种不同的分子，每种都结合一个受体，它们的组合定义了特定的细胞亚群。例如，我最喜欢的细胞是这些具有 CD4 分子的 T 细胞，其中一些还有 CD8 分子。带有 CD8 的是杀伤细胞；它们会去杀死淋巴细胞。我可以在设置门控，然后问：‘CD8 阳性或 CD4 阳性的细胞百分比是多少？’它会生成各种统计分析。这真的是很复杂的软件，因为我还可以改变图形的等高线图——一种不同的表示方式。记住，这里有大约 10 万个事件。

These are the fluorescent molecule colors, if you like. Whatever antibody has that fluorescent molecule will label a particular cell type. I can choose here: 'Change me.' You can see all these are different antibodies labeled differently. It's very complex because there are 20 different molecules, each binding to a receptor, and their combination defines a particular subset of cell. For example, my favorite cells are these T-cells that have the CD4 molecule, and some of them have the CD8 molecule. The ones with CD8 are killer cells; they go and kill the lymphocytes. I can put gates here and say, 'What's the percentage of cells that are CD8 positive or CD4 positive?' It will create all kinds of statistical analysis. This is really complex software because I can change the contour plots of the graphics—a different kind of representation. Remember, there are like 100,000 events here.

而且它运行得这么快，真是不可思议，我没想到它能优化到这种程度。

And it's doing it so fast, which is incredible because I didn't think it would be optimized this way.

而且你完全是用 Codex 构建的这一切？

And you build all of this with Codex?

100% 用 Codex。是的。我花了一些时间，因为有些东西不工作，但尤其是 5.5 版本，我会说：‘好吧，我看不清图表，修一下’，然后写个提示，它就会去工作。所以我做了这个小应用，你可以实际选择你想要的细胞类型。我说我想要一个初始 T 细胞，它就会显示所有我可以选择的标记物。事实上，它甚至会显示，例如，哪些标记物对该细胞类型更相关。这对于开发我们所谓的 panel 非常有用。我在找中央记忆初始细胞、T 细胞和 T87 细胞。如果我选择了这些标记物，那么我就可以回去做我的流式分析。

100% with Codex. Yeah. It took me a while because some things wouldn't work, but especially with 5.5, I would say, 'Okay, I can't see the graphs very well, fix that,' and write a prompt for that, and it would just go and work. So I made this little app where you can actually choose the cell type you want. I say I want a naive T-cell, and it will show me all the markers I can potentially choose. In fact, it will even show me, for example, the ones that are more relevant for that cell type versus another. This becomes extremely useful for developing these what we call panels. I'm looking for the central memory naive cell and a T-cell and T87 cells. If I choose these markers, then I can go back and do my flow analysis.

太棒了。我的意思是，你显然是受过训练的工程师，但构建这些应用肯定要花你几周甚至几个月的时间，对吧？

That's amazing. I mean, you're obviously an engineer by training, but it would have taken you weeks if not months of work to build these apps, right?

我不是软件工程师。我是生物医学工程师，所以我其实不会写代码……我大概花一个月能写个贪吃蛇游戏。做这些应用简直是梦想。所以我做这些小应用来帮助日常工作。

I'm not a software engineer. I'm a biomedical engineer, so I can't really write... I could probably write a snake game if I worked for a month. To make these apps would be kind of like a dream. So I make these little apps that help on a daily basis.

这太不可思议了。关于 T 细胞，再跟我说说。你有没有尝试用更多模型来帮助日常工作，比如用 Codex 甚至像 Imagen 这样的东西？

That's pretty incredible. And on T-cells, tell me more. Have you tried to play with more models to help in your day-to-day work using Codex or even things like Imagen?

是的。我非常有热情的一件事是，我认为 AI 将对生物学产生巨大影响，尤其是能够模拟生物系统，因为它们太复杂了。例如，如果你在造飞机，你不会只说：‘好吧，如果我把这些零件拼在一起，我希望我能得到一架飞机。’你会做模拟——它如何具有空气动力学等等。对于任何复杂的事物，但生物学中我们没有这个，因为部件太多。我的目标是，在某个时候，我们将能够构建所谓的虚拟细胞。我们将能够用 AI 模拟一个完整的免疫细胞，然后是组织，最终是我所说的数字孪生——一个完整的人。我们将需要更多的算力，所以你们最好投资开发这个。但作为开始，我从这样一个应用入手。这基本上是一个受体，叫做 T 细胞受体，位于我刚才展示的 T 细胞表面。它是主要受体，但一个受体极其复杂，取决于它看到的分子亲和力和其他信号。那是决定者，有时会决定你是生是死，因为信号强度可能意味着你患上自身免疫病、清除肿瘤、杀死被病毒感染的细胞，或者造成太多伤害而杀死你。决策点就在那里，受体下面发生着非常复杂的事件。

Yeah. One of the things I'm very passionate about, and I think when I say AI is going to make tremendous impact on especially biology, is to be able to simulate biological systems because they're so complex. For example, if you're building an airplane, you don't just say, 'Okay, if I put these parts together, I hope I'm going to have an airplane.' You do a simulation—how is it going to be aerodynamic and all that. For any kind of complex things, but for biology, we don't have that because there are so many parts. My goal is that at some point we'll be able to build what's called virtual cells. We will be able to mimic a complete immune cell by AI, and eventually tissues, and eventually what I call the digital twin—a complete human being. We're going to need a lot more compute, so you guys better invest in developing that. But to start, I began with an app like this. This is basically one receptor called the T-cell receptor on top of these immune cells, the T-cells I just showed you. It's the main receptor, but one receptor is extremely complex depending on the affinity of the molecules it sees and the other signals it sees. That's the decider that will sometimes decide whether you're going to live or die, because the strength of the signal could mean you get autoimmune disease, you can clear tumors, you kill virally infected cells, or it could cause too much damage and kill you. The decision point is there, and there are very complex events that happen underneath that receptor.

信号通路非常庞大，所以我构建这个模拟器就是为了能够模拟，比如，如果我只用 T 细胞受体，配体有某种质量和剂量，我在这里就能完全控制这些参数。然后我可以让它基于这些参数运行模拟，它会告诉我哪些分子会被激活，哪些不会，甚至还会展示所谓转录因子的磷酸化模式。我还可以说，如果我施加一个压力，比如存在一个抑制性分子，然后我选择另一个信号，会发生什么？它会显示这条通路被阻断，从而产生不同的结果。基本上你可以不断扩展：如果我加入一个小分子来抑制这个分子，输出会怎样？如果我在这里增加更多受体，它们之间会如何相互作用？

There's a huge signaling pathways and so I built this just to be able to simulate for example if I have only the T-cell receptor and then you know if the ligand is a certain quality and the dose is this much I can actually control all of that here and then I can say okay run me a simulation based on that it will run me a simulation it will tell me what molecules are going to get activated which ones will not get activated and it will even show me the phosphorylation patterns of what's called the transcription factors and I can say okay well what if I just put a pressure there's an inhibitory molecule and then I choose a different signal what would happen and then it will show me okay well okay this pathway is stopped now so you're going to get a different type of events and so basically you can just expand this you can say okay what if I just put a small molecule that will inhibit this molecule what will be the output and then what if I just add more receptors to here and how are they going to interact with each other

我很喜欢这一点，因为这不仅仅是可视化数据集或在数据集中搜索，而是一个完整的应用程序，你用它来在浏览器中识别每个特定细胞和场景的输入和输出。

I love that because it's not even just like visualizing data sets or like searching in data sets this is like a complete like application that you build to like identify the inputs and outputs in the browser of every like particular like cell and scenario

确实不可思议。

Incredible yeah

我们看到很多开发者都在用 Codex 和 GPT 5.5 这样的工具来搭建网站、Web 应用和移动应用，但你这个在生物学领域如此复杂和先进。

I mean we see so many builders who are obviously like using tools like Codex and GPT 5.5 to build like websites, web applications, mobile apps, but this is like so intricate and so advanced for the biology field.

是啊，这就是为什么我喝完咖啡就早早起床。我说：‘好，我要把这个加到模拟器里，看看会发生什么。’另一件有趣的事是，Image 2.0 发布后，我彻底爱上了它。你知道，我可以生成极其复杂的图像。比如这张图，我想我也在 X 上分享过。它关于一个免疫图谱，发表在《自然》杂志——全球顶级期刊。我让 Image 基于这个图谱，好玩地给我做一张《自然》封面。它生成了这张令人难以置信的图像，里面所有的免疫细胞都是我过去 30 年研究过的。我当时想：‘天哪，这太精细、太美了。’

Yeah, that's why I wake up early after coffee. I say, 'Okay, so I'm going to add this to the simulator, see what happens.' And then the other thing which was kind of a fun thing to do because after Image 2.0 came out, I just fell in love with it. You know, I can make extremely complex images. For example, this image, I think I shared it on X as well. This was about an immune atlas that was published in this journal called Nature, the top journal in the world. And so I basically asked Image to said, you know, create me a cover page of Nature just for fun based on this atlas. And it made this unbelievable image where you can see all these immune cells are the ones that I've actually worked for the past 30 years. So I was like, 'Oh my god, you know, this is so detailed and so beautiful.'

然后我说：‘如果我把这张图放进 Codex，让它做成一个交互式网站，让它活起来呢？’我就这么做了。Codex 生成了这些漂亮的动画，展示了细胞之间如何相互作用。我可以点击它们，它会显示这是记忆细胞，这是效应细胞，并提供各种信息。比如，如果我阻断 PD1 这个分子——它在癌症治疗中非常重要，阻断它可以治愈很多癌症，这就是所谓的检查点抑制剂。这本来完全是个好玩的事，但我后来意识到，哇，我甚至可以改变参数，比如根据你的年龄，免疫细胞的行为会如何变化。这就像一个小型模拟器。而这一切都始于一张图片。

And then I said, 'What if I just take that image and put it into Codex and tell it to sort of make it live in an interactive website?' And that's what I did. And then Codex just came up with these beautiful animations and how these cells are interacting with each other. I can click on them. It shows me what you know this is a memory cell. This is an effector cell. It gives me all kinds of information about them. You know if I block it for example this molecule PD1 is very important in cancer therapy. If you block this molecule you can cure a lot of cancers called checkpoint inhibitor. And you know this was a totally a fun thing to do but then I realized wow I can even change for example based on your age how your immune cells are going to behave. This is like a mini simulator here. It all started with just one image.

就这一张由 GPT 生成的图片。

Just this one image GPT image to generate.

没错。我没有说做这个或做那个，我只是说：我希望你理解这张图片，并创建一个交互式模拟器网站。就这么简单，Codex 就帮我建好了。

Exactly. And I didn't say okay do this or do that. I said I want you to understand this image and create an interactive simulator website. That's all I did. And Codex just built this for me.

太棒了。是啊，它也是一个很好的教育工具，对吧？因为如果你想深入学习某个主题，了解那个主题。

Amazing. Yeah, it's also such a great tool for education as well, right? Because like if you want to learn into like dig into any topic, learn about that topic.

对。

Yeah.

你现在不仅可以生成图片，还可以为那个主题生成模拟器。这太不可思议了。听起来你构建了很多应用。感谢你的分享，这让我得以一窥你的工作、你的思考方式，以及你使用 Codex 的方式——我至今还没用 Codex 构建过这些应用，所以这对我来说非常吸引人。

You can now generate image but also just generate like simulators for that topic. It's pretty incredible. Sounds like you're building so many apps. Like thank you for sharing. This is like a window into your work and how you think and how you use Codex in a way that like I've not used Codex to build these apps so far. So this is fascinating for me.

我再展示一个我认为非常有用的应用。作为生物医学工程师，我们想要操控细胞。在某种程度上，细胞就像编程好的代码软件，我希望有一天我们能有生物学版的 Codex，可以完全编程细胞。事实上，我们已经开始用这些技术来基因编辑细胞。我实际上在 25 到 30 年前就开发了其中一些技术，现在我们有了 CRISPR。我相信很多人都听说过。基本上，CRISPR 可以靶向任何基因，改变固定突变，删除基因，或使其过表达。这就像基因组工程。但问题在于，这非常复杂。比如你有一个 2000 个核苷酸的基因，你要靶向哪里呢？

Let me just show one more app which I think is very useful. As a biomedical engineer, what we want to do is we want to manipulate the cells. In a way, the cells are kind of like programmed code software, and I'm hoping that one day we'll have a Codex for biology where we can totally program the cells. In fact, we started doing that with these technologies to genetically edit cells. I actually developed some of that 25 years ago, 30 years ago, and now we have this thing called CRISPR. I'm sure a lot of people have heard of it. Basically, CRISPR is you can target any gene, you can change fixed mutations, you can delete the gene, you can overexpress it. It's like a genome engineering. But the problem is that again this is very complex. So you have a gene that's let's say 2,000 nucleotides, where are you going to target, right?

所以你需要某种计算来判断它是否具有特异性，是否会产生这种效果。已经有类似的工具了。但我说，我想要自己的工具，这样我就能……所以我构建了这个应用，你可以选择任何你想要的基因。比如，这是 CD4 基因，我之前提到过它位于 T 细胞表面。它会立即从数据库中拉取 CD4 的 DNA 序列。

Uh so you have to have some sort of a computation to say that that's going to be specific, it's going to do this effect or not. And there are some tools like that. I said well I want my own you know so that I can so I built this application where you can actually choose any gene you want. So let's say this is a CD4 gene which I mentioned about that on the surface of T-cells. It immediately pulls the sequence from the database the DNA of the CD4.

是的，没错。然后它会给出所有潜在的靶点。这些是 20-22 个核苷酸的区域，但这是一个非常大的基因。它会排序，说你应该选这个，因为这个很好。我可以说，好，把这个加进去，那个加进去，然后复制，交给在线公司，他们会合成并寄给我，我就可以进行工程改造和实验。我可以可视化它，它会显示所有被靶向的区域。这里有一些其他应用没有的功能，这就是它的美妙之处。比如，我说：你能为我创建一个文库吗？如果我有多个基因，想要很多不同的 CRISPR 靶点呢？所以我构建了这个版本，我只需在这里输入基因名称，然后点击设计文库，它就会为我生成不同的 CRISPR 分子。

Yeah, exactly. And then it gives me all the potential targets. So these are 20-22 nucleotide regions but it's a very big gene. And then it ranks them says okay well you should probably choose this one because this is good. And I can just say okay add that there, add that there and then I can copy it and then give it to an online company and they will synthesize that and send it to me and I will engineer and do the experiment. I can visualize that. It shows me all the regions where these are being targeted. So there are some features here that are not present in other applications. That's the beauty. For example, I said okay can you create a library for me? So what if I have multiple genes and I want many different CRISPR targets. So I built this version where I can just write the genes name here and all I do is just click on design library. It will make different CRISPR molecules for me.

太棒了。我们现在看到的是你 Mac 上的原生应用，对吧？

That's amazing. And this is a native app that we're looking at on your Mac, right?

是的，用 Swift 构建的原生 Mac OS 应用。我还会做一个 iPad 版本，这样我就可以带进实验室了。

Native Mac OS app that's built with Swift. And I actually I'm going to make an iPad version too so that I can carry inside the lab.

非常感谢你的分享。能了解你工作的幕后，真是太棒了。

Well, thank you so much for sharing that. This is incredible to see the behind the scenes of your work.

是啊。现在你明白为什么我喝完咖啡后就……

Yeah. Now you understand why after coffee I...

我现在完全理解你的日常了。这对我来说更有意义了。

I totally understand the routine now. It makes so much more sense to me.

你之前提到的一个点我想深入聊聊，就是数字孪生这个概念。以你在免疫学、癌症、生物学方面的视角，你觉得数字孪生什么时候能变得可行？

One thing that you mentioned earlier that I'd love to double click on is this idea of like a digital twin, right? And as you have like your vantage point in like immunology, cancer, biology, when do you think this idea of a digital twin becomes viable?

首先，为什么需要数字孪生，为什么需要 AI？我所说的数字孪生，是指一个人完整的生物系统。我们的生物学不仅仅是你从外部测量的那些，还有免疫系统、代谢物、肠道里数万亿的细菌、激素等等，是一个非常非常复杂的系统。当然，还有你的基因。基因和环境几乎决定了一切：你会不会生病、什么时候生病、对某种治疗有没有反应。我们能不能在疾病发生前就预测？而且这必须高度个性化。实际上，你要治疗的是病人，而不是疾病。但因为生物学太复杂，我们一直给患有同一种疾病的数百万人用同一种药。比如他汀类药物，数百万人都在用，但只对一小部分人有效。所以，我们能不能达到这样一个点：AI 能完全模拟你的生物学，包括你的基因组、代谢物、蛋白质、免疫系统？如果能做到，我们就可以开始问：如果我在健康上做这个改变会怎样？如果我给他用这种药会怎样？也许我会根据你的生物学，按照 AI 告诉我们的，为你量身定制治疗方案。这样我们就进入了完全个性化的阶段，药物将接近 100% 有效、0% 副作用。这意味着临床试验将从 5 到 10 年缩短到可能 5 到 10 天。AI 将真正为你做临床试验。这就是令人难以置信的加速，因为我说在未来十年左右，我们将能治疗所有疾病。最终在 15 年内，我们将逆转衰老，人们可以活几百年。有人说：“这听起来很疯狂，是科幻小说。你知道，我们花了 50 年才在癌症治疗上取得一点进展。”他们没有算到的是，AI 正在指数级进步，而且这还假设算力会大幅增加，因为即使把全世界所有的计算机加起来，也不足以模拟生物系统。有数万亿个组件。所以如果我们达到那个点——我相信在未来 5 到 10 年内就能做到，并且我们有了超级智能——那么基本上我们就可以用 AI 模拟这个数字孪生，然后不是在人身上做实验，而是用 AI 针对你的生物学做实验，这将改变医学，改变一切。

First of all, why digital twin and why we need AI? The reason we need digital twin and when I say digital twin I mean a person's complete biological system. So our biology is not just what you measure outside as I showed you. There's the immune system, there's our metabolites, there's our trillions of bacteria in our gut and hormones and so on so forth. It's very very complex system. And of course, there's your genetics. And your genetics and your environment decide pretty much everything. Whether you're going to get sick or when are you going to get sick, whether you're going to respond to a given treatment. Can we predict before the disease happens? And it has to be very personalized. In fact, you have to treat the patient, not the disease. But because the biology is so complex, we have been giving the same drug with the same disease to millions of people. For example, statins are used by millions of people, but it only affects a small portion of them. So can we come to a point where AI can completely simulate the biology and your biology with your genome with your metabolites with your proteins your immune system and if you can get that point then we can start to ask this question what if I make this change in my health or what if I give them this drug or maybe I will basically order a treatment just for you based on your biology, what the AI is going to tell us. And so, we're getting into a point of complete personalization where the drugs will have nearly 100% effect with 0% side effect. And that means that the clinical trials that we do are going to be accelerated from 5 10 years to maybe 5 10 days. Literally AI is going to do the clinical trial for you. And that's the incredible acceleration because when I say in the next decade or so, we're going to treat all of the diseases. Eventually within 15 years, we're going to reverse aging and people will be able to live hundreds of years. People say, "Well, that sounds crazy. That's science fiction. You know, it took us 50 years for cancer just to be able to treat a little bit." What they don't calculate is that AI is advancing exponentially and this is also assuming that the compute is going to greatly increase because even the current available compute if you put all of the computer in the world is not enough to simulate biological systems. There are trillions of components. So if we get to that point which I believe we'll be able to do in the next 5 10 years and we'll have the super intelligence then basically we simulate this digital twin with AI and then we do an experiment not on humans but with AI for your biology and that will change medicine that will change everything.

那么举个例子，比如一个癌症患者，如果我们有他的数字孪生，医生能做什么和今天不同的事情？他们基本上就是在那个数字孪生上尝试不同的假设和实验，看看它如何反应，然后我们就这么治疗稳定组和治疗组，差不多是这样吗？

So if we take the example for instance of like a cancer patient say and if we had their digital twin what could a doctor do differently versus what they do today would they just basically try different hypothesis and experiments on that digital twin and see how that reacts and that's how we treat the stable group and the treatment group almost

完全正确。实际上，癌症和肿瘤学是今天最接近个性化的领域，因为即使是同一种癌症类型也有许多不同的突变。所以如果你是一个肺癌患者，你的肿瘤医生要做的第一件事可能就是测序那些突变的基因，因为根据哪个基因突变，有针对性的不同药物。例如，1% 的肺癌可以用一种特定的药物，那种药对那 1% 非常有效，但对 99% 无效。这些都是公司正在开发的智能药物。所以我们可以达到这样一个点：我们可以为你的所有突变创造一种药物。事实上，澳大利亚有一个案例，一位计算机科学家用 ChatGPT 和 Grok 为他的狗创建了一种 RNA 疫苗。那是最个性化的了，因为那个 RNA 疫苗就是专门针对那只狗的癌症突变创建的，而且有很多相关试验在进行。所以我认为这将彻底改变，因为正如我所说，不仅基于你的基因和突变，还有你的免疫系统。例如，免疫系统在杀死癌细胞方面非常有效，这已经是一场革命，我们称之为免疫疗法，但它不是对每个人都有效。为什么？为什么有些人的免疫细胞能够识别并杀死癌症，而其他人不能？有些会变得耗竭等等。所以如果我们能弄清楚这一点，当然还有副作用，因为免疫系统非常危险，如果过度激活，实际上会造成很大伤害。

Exactly. Actually cancer and oncology is as close as we are today for personalization because even the same cancer type have many different mutations. And so you probably if you were a lung cancer patient the first thing your doctor oncologist would do is to sequence the genes that are mutated because depending on which gene is mutated you have a different drug for that. For example 1% of lung cancers can use a particular drug and that drug is very effective in that 1% but it's not effective at the 99%. So these are smart drugs that companies are developing. So we can come to a point where we can literally create a drug for all the mutations that you have. In fact there was a case from Australia this computer scientist using ChatGPT and Grok you know created an RNA vaccine for his dog right. So that's as personalized as you can get because it was that RNA vaccine was created just for that dog's cancerous mutations and there's a lot of trials going on with that so I think it will change completely because as I said based on not only your genes and mutations but your immune system for example immune system is extremely effective in killing cancer cells and this has been a revolution what we call immunotherapy but it doesn't work in everybody. Why is that? Why is some people's immune cells able to recognize the cancer and kill but others don't? Some of them become exhausted and whatnot. So if we could figure that out and of course there are also side effects because immune system is very dangerous if you activate it too much it can actually cause a lot of harm.

我很好奇，你看起来对 AI 深信不疑、接受得这么快，同时又是一名医学博士，你周围的人对 AI 怎么看？你会试着让他们也像你一样快速采用这些工具吗？

I'm curious like you seem so AI-pilled and so quick to adopt while also being like an MD like what are the people around you thinking about AI? Like do you try to get them also to adopt these tools as fast as you are?

是的，我非常努力地尝试，我觉得他们认为我完全疯了。不过我想现在他们开始看到潜力了，而且我从 ChatGPT 3.5 出来就开始这么说。我认为人们非常犹豫，我能理解，因为这是全新的东西，人类思维无法理解这种指数级的进步。例如，人们用一年半前的 GPT-4o，这在 AI 世界里已经是远古时代了，然后说它幻觉太多，回答得不好。实际上，即使是 GPT-5.4 和 5.5 之间的差异也是天壤之别。如果你不断实验，然后假设它会越来越好。我信任 AI 模型，甚至对于我研究了 30 年的领域也是如此，因为最近 GPT-5.5 Pro 模型给了我一份报告，我几乎哭了。我说这怎么可能？

Yeah, I try very hard and I think they think that I'm totally crazy. Although I think nowadays they're starting to see the potential and I've been saying this since ChatGPT 3.5 came out. I think people are very hesitant and I can understand that because this is something so new that human mind is not able to comprehend this exponential advance. For example, people use GPT-4o a year and a half ago, which is like ages in AI world, and say, well, you know, it was hallucinating too much. It didn't really answer to me very well. In fact, even the difference between GPT-5.4 and 5.5 is day and night. If you constantly experiment and then assume that it's going to get better and better. I trust AI models even for things that I've been working on for 30 years because recently GPT-5.5 Pro model gave me this report that I almost cried. I said how is this possible?

我记得在网上看到过你关于那次经历的帖子。

I remember seeing your post online about that experience.

没错。因为那个最新模型跨过了一个门槛。当然，GPT-5 Pro 和之前的 5.4 在理解知识和模式、把它们整合起来方面已经很棒了。但 5.5 所做的，几乎就像它拥有我在实验室工作 30 年的相同经验，因为有些事情是直觉，你就是知道，它不在文献里，你能感觉到。比如我会和学生打赌，我说做这个实验，我赌它会这样起作用，他们有时会赌，但每次都输，因为你知道，我就是知道它会起作用。

Exactly. Because that latest model crossed the threshold. Of course GPT-5 Pro and 5.4 before they were fantastic understanding the knowledge and the patterns and putting them together. But what 5.5 did was almost as if it had the same experience that I had working in the lab for 30 years because there are certain things are intuition you just know it's not out there in the literature you feel that for example I would bet with my students I say do this experiment I'll bet you it's going to work this way and they would bet sometimes and they would lose 100% of the time because you know I just know that it's going to work.

它经历了那么多迭代，就像你的直觉一样。

It had so many reps that like your intuition.

是的。不知何故，你会有那种直觉，而这就是我对 5.5 的感受——它预测了我们做过的一个非常复杂的实验的结果，准确率达到了 100%，这太不可思议了。

Yeah. Somehow that you have that intuition and that's what I felt with 5.5, that it predicted the outcome of an experiment, a very complex experiment we had done, at 100% level, which was incredible.

我很好奇，如果我们继续以你这边所经历的速度进步，你想象几年后你的日常工作会是什么样子？对于像你这样的人和你周围的生物学研究人员来说，什么会变得截然不同？

I'm curious, if we continue on this rate of progress that you're also experiencing on your side, what do you imagine your day-to-day looking like in a few years from now? What will be unmistakably different for people like yourself and researchers around you in this world of biology?

嗯，再说一次，我说的一些话可能听起来很激进，但对我来说幸运的是——也许对那些不相信这一点的人来说不幸的是——将会发生一个根本性的转变。事实上，我称之为科学 2.0 或 3.0。我们做科学的方式将彻底改变。过去我们提出想法、花数周规划实验、再花数月分析结果的日子已经结束了。学生或其他科学家必须意识到，我们正处于一个极度加速的时间尺度上。我们做科学的方式将是一群 AI 智能体提出假设。它们已经能够提出假设，因为你能想到的东西几乎是无穷无尽的。你怎么知道这个假设比那个更好？你会得到最好的想法。然后它会为你模拟实验。我可以做一千个实验，但哪个会成功？我不知道。如果它能引导我到一个点，说‘好的，这 10 个实验因为这样那样的原因更可能成功’，那么我就可以专注于那些。那个实验成功的几率会大大增加。我得到的数据几乎是即时反馈。我会把它交给其他 AI 智能体，它们会立即分析并反馈给主智能体，而那个主智能体会提出新的假设并规划新的实验。所以我想我的角色就是告诉智能体：‘去搞清楚这个。这是你应该走的方向。我想治愈肺癌，朝那个方向努力。’

Well, again, some of the things I say may sound very radical, but fortunately for me, but maybe unfortunately for those who don't believe in this, there's going to be a complete fundamental shift. In fact, I call this Science 2.0 or 3.0. The way we do science is going to completely change. The days where we would come up with an idea, plan experiments for weeks, and then analyze them for months, those days are over. Students or other scientists have to realize that we are in a very accelerated time scale. The way we do science is going to be a bunch of AI agents that will come up with the hypothesis. They're already able to come up with the hypothesis because the number of things you can come up with are almost endless. How do you know that hypothesis is better than that one? You'll get the best ideas. Then it will simulate the experiments for you. I can do a thousand experiments, but which one's going to work? I don't know. If it could drive me to a point saying, 'Okay, these 10 experiments are likely to work more because of this and this,' then I can just focus on that. The chances of that experiment working will increase tremendously. The data I get, I have almost instant response. I'll give it to other AI agents, and they will immediately analyze and feed it back to the master, and that master is going to make a new hypothesis and plan for a new experiment. So I guess my role is going to be just telling the agents, 'Go figure this out. This is the direction you should go. I want to cure lung cancer, work on that.'

去探索吧。

Go explore.

完全正确。然后还是得有人做实验，但我认为实验室会实现自动化。这已经在发生了。将会有大量机器人做很多实验。人们说：‘我还会有一份工作吗？’有个概念叫杰文斯悖论。如果我们能做得这么多，那么我们就能做得更多，尤其是在生物学领域。我们可以达到生物工程的程度，但我们现在只了解大约 10% 的生物学。想象一下我们还能学多少、学得多快。一旦我们有了那种能力——就像我现在构建应用一样——我将能够构建新的细胞类型、新的组织。可能会有成千上万的生物工程师坐在电脑前进行模拟和构建。这不仅会从根本上改变生物学，我认为每一个领域：物理学、材料科学、化学。在药物发现方面，你已经可以在几小时内完成过去需要数年的事情。还有为医生治疗病人进行的临床试验。整个堆栈都将改变并加速。

Exactly. And then somebody still has to do the experiment, but I think the labs are going to get automated. This is already happening. There's going to be a bunch of robots doing lots of experiments. People say, 'Will I still have a job?' There's something called Jevons paradox. If we are able to do so much, then we will be able to do so much more, especially in biology. We can get to a point of bioengineering, but we only know like 10% of biology. Imagine how much more, how fast we can learn. Once we have that power, like I'm building apps now, I'll be able to build new cell types, new tissues. There will be maybe thousands of bioengineers in front of their computers simulating and building. This is going to fundamentally change not just biology, I think every field: physics, material sciences, chemistry. Already for drug discovery, you can do it in a matter of hours that took years to do. And doing clinical trials for doctors treating patients. All of that stack is going to change and accelerate.

说到这些其他领域，对于不像你那样从事科学、但可能在不同领域的人，关于你的经历你会给他们什么建议？你在这方面已经很久了，一直在思考 AI，并且每一步都怀着用 AI 构建的好奇心。你会建议他们如何重新思考自己的领域和工作？

Speaking of these other fields, for people who are not in science like you are, but maybe in different fields, what would you advise them about your journey? You've been at this for a long time, thinking about AI and having this curiosity to build with AI every step of the way. What would you advise them in terms of how they have to rethink their own field and their own work?

是的，我有一个优势：我的工作、我的生活依赖于不断做实验。这非常痛苦，因为尤其是在生物学中，你做的实验有 95% 到 98% 会失败。所以我很习惯失败。这就是为什么它叫实验。你说：‘好的，我们试试这个，试试那个。’然后你培养出一定的韧性、主动性和好奇心去尝试。这就是为什么我对 Codex 如此兴奋。我向你展示了那些成功的东西，但在此之前有很多失败。我做的很多应用效果不太好，结果也不理想。但你不应该放弃。我给人们的建议是：在 AI 时代，真正重要的是主动性和好奇心。不要害怕。不断尝试，不断用 AI 做实验。说：‘如果我能这样做呢？’因为你现在可以问那个‘如果’的问题了。以前，问那个问题太昂贵了。假设你有一个公司网站。你可能花了数千美元制作它，然后它并不完美，但人们说：‘好的，够好了。我不想再动它了。’现在，你可以说：‘如果我能改一下这个，做成这样呢？’只需几分钟，你就有了一个新网站或一个新产品，你可以设计它，然后 3D 打印出来，例如。我认为一切都可以应用，但你必须要有尝试的勇气，因为实验的成本如此之低。你为什么不这样做呢？这些都是非常简单的日常事情，但我认为你可以真正将其扩展到生活中的几乎任何事情。你只需要拥抱它，把 AI 看作是非常积极的东西。我看到很多负面情绪：‘AI 会做这个，AI 会做那个。’我相信恰恰相反。它会真正将我们提升到一个黄金时代。我真的对此非常兴奋。顺便说一句，我认为 AI 研究人员都是英雄，因为这将是人类最伟大的变革。我对未来超级兴奋。

Yeah, so I have one advantage: my job, my work, my life depends on constantly doing experiments. This is very painful because especially in biology, 95 to 98% of the experiments you do fail. So I'm very used to failure. That's why it's called experiment. You say, 'Okay, we'll try this, try that.' And you develop a certain resilience and agency and curiosity just to try. That's how I was so excited about Codex. I showed you the things that worked, but there were a lot of failures before that. Many apps that I did didn't work very well and didn't come out very well. But you shouldn't give up. My advice to people would be: in the age of AI, really agency and curiosity are the only things that matter. Don't be scared. Constantly try, constantly experiment with AI. Say, 'What if I can do it this way?' Because you can now ask that 'what if' question. Before, it was just too expensive to ask that. Let's say you have a website for your company. You probably spent thousands of dollars to have it made, and then it wasn't perfect, but they say, 'Okay, it's good enough. I don't want to touch it.' Now, you can say, 'What if I can just change this and make it this way?' In just a matter of minutes, you have a new website or a new product that you can design and then 3D print it, for example. I think everything can be applied, but then you have to have that courage to experiment because the cost of experimentation is so low. Why shouldn't you do that? These are very simple daily things, but I think you can really extend that to pretty much anything in your life. You just have to embrace it and look at AI as something incredibly positive. I see a lot of negativity: 'AI is gonna do this, AI is gonna do that.' I believe just the opposite. It's going to really enhance us to a golden age. I'm really super excited about it. I think AI researchers are all heroes, by the way, because this is going to be the greatest transformation of humanity. I'm super excited for the future.

非常感谢你，Derya。感觉太棒了。以非常积极的方式结束对话。我们迫不及待想看看你接下来会做什么，以及你将如何推动 Codex 和这些前沿模型，将整个世界融合在一起，并致力于这个数字孪生的想法。

Thank you so much, Derya. Feels like it's amazing. A way to close the conversation very positive. We can't wait to see what you're going to do next and how you're going to push Codex and these frontier models to bring all of this world together and work on this digital twin idea.

当然。很乐意再来一期节目。

Absolutely. Happy to come back for another episode.

我们很希望几个月后你能回来，看看你取得了多少进展。在那之前，享受你在加州的时光。

We'd love to have you back in a few months to see how much progress you've made. Until then, enjoy your time in California.

谢谢。非常感谢。这非常有趣。

Thank you. Thank you so much. This was a lot of fun.